JP2007299071A - Reputation information processing system, reputation information processing method, and reputation information processing program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a reputation information processing system, a reputation information processing method, and a reputation information processing program by which reputation information useful for a specific user can be acquired. <P>SOLUTION: A PC 100 as a reputation information processing system comprises a usefulness acquisition part 10 for acquiring reputation information showing the evaluation of an object and usefulness of the reputation information for a specific user, a morphological analysis part 20 for executing the morphological analysis of the reputation information, a classification model generation part 30 for generating a classification model consisting of morphemes obtained by the morphological analysis of the reputation information corresponding to the usefulness for the specific user, and a classification part 40 for classifying the reputation information based on the classification model according to the usefulness for the specific user. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention is executed in a reputation information processing system that performs processing for classifying reputation information representing evaluations of various objects such as movies and books, a processing method in the reputation information processing system, and the reputation information processing system. Related to the program.

  There is a technique to classify reputation information, which is a subjective evaluation of opinions and opinions about various objects such as movies and books, according to the evaluation object and the content of the evaluation (positive evaluation or negative evaluation). It is known (see, for example, Patent Document 1).

Reputation information classification generally holds in advance a dictionary that defines adjectives, adjective verbs, words that are positive or negative expressions (morphemes) such as function words, and evaluation values that represent the degree of evaluation, This is realized by calculating the total evaluation value in sentence units or document units. For example, a dictionary defined as “2” for “interesting”, “1” for “like” and “−1” for “bad” is retained, and reputation information is “that work cast is bad, but the story is interesting” In the case of a sentence “I like”, the evaluation value of the reputation information is −1 + 2 + 1 = 2 and is a positive numerical value, and therefore, it is determined that the reputation information is affirmative. In calculating the evaluation value, syntax information may be used in addition to morpheme information (see, for example, Non-Patent Document 1).
Japanese Patent No. 2402599 Matsumoto S, Takamura H, Okumura M, “Sentiment Classification using Word Sub-Sequences and Dependency Sub-Trees.”, In Proceedings of the 9th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD-05), p. 301-310, 2005

  However, the above-described conventional method focuses on grasping the overall tendency of reputation information, and it is difficult to say that sufficient classification is made when using each piece of reputation information. In particular, when a certain user uses individual reputation information, it may be unclear whether the reputation information is useful for the user.

  For example, in the conventional method described above, the reputation information is “interesting” and “It ’s a stupid story when you think calmly, but Hitomi Shiraki ’s celebs are crazy, Junichi Okamoto, Because both Matsuki Jun's screens are handsome and supportive, it has become a mysterious and interesting movie. ”Both of the reputation information include positive expression morpheme that is“ interesting ”. Treated as reputation information.

  However, when a user uses reputation information as a reference when watching a movie, for example, it is unclear whether the reputation information “interesting” is interesting for the user who wrote the reputation information. It is difficult to use. In such a case, if the user is familiar with the nature and preferences of the writer, the user can determine to a certain extent whether something interesting for the writer of reputation information is interesting for himself, but reputation information from the Internet etc. When using, reputation information is created by a large number of unspecified number of writers, so such a determination is difficult.

  An object of the present invention is to solve the above-described problems, and provide a reputation information processing system, a reputation information processing method, and a reputation information processing program that can obtain reputation information useful for a specific user.

  The reputation information processing system according to the present invention includes reputation information representing an evaluation of an object, acquisition means for acquiring the usefulness of the reputation information for a specific user, and syntactic / semantic analysis of the acquired reputation information. It comprises analysis means for performing morpheme analysis, and generation means for generating a classification model composed of morphemes obtained by the morpheme analysis of reputation information corresponding to the acquired usefulness for a specific user.

  According to this configuration, the reputation information and the usefulness of the reputation information for the specific user are acquired, and the morphological analysis is performed on the reputation information. The usefulness for the specific user and the morpheme obtained by the morphological analysis Based on the above, a classification model is generated. Since this classification model includes morphemes obtained by morphological analysis of reputation information corresponding to the usefulness for a specific user, a large amount of reputation information can be converted into usefulness for a specific user by using the classification model. It becomes possible to classify accordingly. Here, the usefulness is, for example, a numerical value representing the degree of usefulness of reputation information for a specific user.

  In the reputation information processing system according to the present invention, the reputation information may be sentence-by-sentence information, and the usefulness may be given to each sentence.

  In the reputation information processing system of the present invention, the generation means includes a set of morphemes included in a higher predetermined number of appearance frequencies in reputation information whose usefulness is a predetermined value or more among morphemes obtained by the morpheme analysis. And a morpheme set included in the upper predetermined number of appearance frequencies in the reputation information whose usefulness is less than a predetermined value may be generated.

  A morpheme obtained by morpheme analysis is highly likely to be a morpheme that should be included in reputation information with a high degree of usefulness if the frequency of appearance in reputation information with a usefulness equal to or greater than a predetermined value is high. If the appearance frequency in reputation information that is less than a predetermined value is high, it is highly likely that the morpheme should be included in reputation information with low usefulness. Therefore, a classification having a set of morphemes with higher appearance frequency in reputation information having usefulness equal to or higher than a predetermined value and a set of morphemes with higher appearance frequency in reputation information having usefulness less than a predetermined value. By generating a model, an appropriate classification model can be generated from the viewpoint of the appearance frequency of morphemes in reputation information.

  From the same viewpoint, the reputation information processing system according to the present invention includes the generation unit included in a higher predetermined number of appearance frequencies in reputation information whose usefulness is a predetermined value or more among morphemes obtained by the morpheme analysis. And a set of morphemes included in a lower predetermined number of appearance frequencies in reputation information whose usefulness is less than a predetermined value, and included in an upper predetermined number of appearance frequencies in reputation information whose usefulness is less than a predetermined value. In addition, a classification model having a set of morphemes included in a lower predetermined number of appearance frequencies in reputation information whose usefulness is greater than or equal to a predetermined value may be generated.

  According to this configuration, the morpheme obtained by morpheme analysis may be a morpheme that should be included in reputation information with high usefulness if the frequency of appearance in reputation information with a usefulness less than a predetermined value is low. Appropriate classification considering that it is highly likely that it is a morpheme that should be included in reputation information with low usefulness if the frequency of appearance is low in reputation information that is high and usefulness is greater than or equal to a predetermined value A model can be generated.

  In the reputation information processing system of the present invention, the acquisition unit may acquire the usefulness of the reputation information for the specific user according to the setting of the usefulness for the reputation information by the specific user. .

  According to this configuration, since the usefulness of the reputation information is acquired according to the setting made by the specific user, the generation of the classification model more accurately corresponding to the usefulness for the specific user, and further, the classification of the reputation information can be performed. It becomes possible.

  In the reputation information processing system of the present invention, the acquisition unit may acquire the usefulness for the specific user for a part of the reputation information.

  In addition, the reputation information processing system of the present invention may have classification means for classifying given reputation information according to the usefulness for the specific user based on the classification model.

  Further, the reputation information processing system of the present invention is configured so that the acquisition means stores the reputation information until the number of all pieces of reputation information of the reputation information whose usefulness for the specific user is acquired reaches a first predetermined value. You may make it continue acquisition of the usefulness for a specific user.

  According to this configuration, it is possible to minimize the acquisition of the usefulness of reputation information to the minimum necessary for the classification model used for reputation information classification to be appropriate, and to quickly generate a classification model. As well as being possible, it is possible to reduce the processing load and the load of the usefulness setting operation by a specific user.

  From the same point of view, in the reputation information processing system of the present invention, the number of pieces of reputation information whose usefulness for the specific user is acquired by the acquisition unit according to the setting by the specific user is a second predetermined value. Until then, acquisition of the usefulness for the specific user of the reputation information according to the setting by the specific user is continued, and the generation means acquires the usefulness for the specific user according to the setting by the specific user When the number of reputation information reaches the second predetermined value, a classification model composed of morphemes obtained by the morpheme analysis of reputation information corresponding to the acquired usefulness for a specific user may be generated. .

  The reputation information processing method according to the present invention includes reputation information representing an evaluation of a target, an acquisition step of acquiring the usefulness of the reputation information for a specific user, and at least syntactic and semantic analysis of the acquired reputation information An analysis step for performing morpheme analysis and a generation step for generating a classification model composed of morphemes obtained by the morpheme analysis of reputation information corresponding to the acquired usefulness for a specific user are provided.

  Further, in the reputation information processing method of the present invention, the generation step includes a set of morphemes included in a higher predetermined number of appearance frequencies in reputation information whose usefulness is a predetermined value or more among morphemes obtained by the morpheme analysis. And a morpheme set included in the upper predetermined number of appearance frequencies in the reputation information whose usefulness is less than a predetermined value may be generated.

  The reputation information processing program according to the present invention includes reputation information representing an evaluation of a target, an acquisition step of acquiring the usefulness of the reputation information for a specific user, and at least of syntactic and semantic analysis on the acquired reputation information An analysis step for performing morpheme analysis and a generation step for generating a classification model composed of morphemes obtained by the morpheme analysis of reputation information corresponding to the acquired usefulness for a specific user are provided.

  In the reputation information processing program of the present invention, the generation step includes a set of morphemes included in a higher predetermined number of appearance frequencies in reputation information whose usefulness is a predetermined value or more among morphemes obtained by the morpheme analysis. And a morpheme set included in the upper predetermined number of appearance frequencies in the reputation information whose usefulness is less than a predetermined value may be generated.

  According to the present invention, a classification model is generated based on the usefulness of reputation information for a specific user and a morpheme obtained by morpheme analysis. Based on the classification model, a large amount of reputation information is converted into a large amount of reputation information. Information can be classified according to the degree of usefulness for a specific user, and reputation information useful for the specific user can be obtained.

  Embodiments of the present invention will be specifically described with reference to the drawings. FIG. 1 is a diagram showing a hardware configuration of a personal computer (PC) as a reputation information processing system. A PC 100 as a reputation information processing system shown in FIG. 1 includes a CPU 101, a memory 102, a hard disk drive (HDD) 103, an operation unit 105, a monitor 106, and a communication unit 108 connected to an internal bus 107.

  FIG. 2 is a functional block diagram of the PC 100 as a reputation information processing system. The PC 100 shown in FIG. 2 classifies reputation information representing evaluations of various objects such as movies and books, and includes a usefulness acquisition unit 10 corresponding to an acquisition unit and a morpheme analysis unit 20 corresponding to an analysis unit. And a classification model generation unit 30 corresponding to the generation unit and a classification unit 40 corresponding to the classification unit. In the hardware configuration of FIG. 1, each of these functional blocks is realized by the CPU 101 executing a predetermined program read from the HDD 103 and stored in the memory 102 in response to an operation of the operation unit 105.

  The usefulness acquisition unit 10 inputs reputation information, and adds tag information (usefulness tag information) according to the usefulness for the specific user according to an operation instruction of the specific user for a part of the reputation information.

  The morpheme analysis unit 20 performs morpheme analysis of reputation information to which usefulness tag information is added. Based on the usefulness tag information added to the reputation information and the morpheme obtained by the morpheme analysis of the reputation information, the classification model generation unit 30 includes the morpheme to be included in the highly useful reputation information, the usefulness level A reputation information classification model including morphemes to be included in low reputation information is generated. Based on this classification model, the classification unit 40 classifies reputation information into those that are useful for a specific user and those that are not useful.

  Hereinafter, the operation of the PC 100 as the reputation information processing system will be described with reference to a flowchart. FIG. 3 is a flowchart showing the operation of the PC 100 as the reputation information processing system.

  The usefulness acquisition unit 10 uses an arbitrary search engine or the like and inputs reputation information to be classified from various sites on the Internet, for example, via the communication unit 108 (S101).

  Next, the usefulness acquisition unit 10 randomly extracts a part of the reputation information, and displays the extracted reputation information (first reputation information) and a screen that prompts the user to set the usefulness of the first reputation information. Are displayed on the monitor 106. The specific user operates the operation unit 105 to set the usefulness level representing the degree of usefulness for each first reputation information. The usefulness acquisition unit 10 adds the tag information of the usefulness set by the specific user to the corresponding first reputation information (S102). The first reputation information to which the usefulness tag information is added and the other reputation information are sent to the morphological analysis unit 20. Note that the number of pieces of first reputation information to be extracted and the range of values that the usefulness can take depend on the learning model used in the classification model generation unit 30 described later.

  Figure 4 shows SVM (Non-Patent Document Joachims, T, “Text Categorization with Support Vector Machines: Learning with Many Relevant Features.”, In Proceedings of the European Conference on Machine Learning. , Springer, 1998)) is a diagram showing an example of setting the usefulness.

  Again, referring back to FIG. The morpheme analysis unit 20 performs morpheme analysis of the first reputation information to which the usefulness tag information is added, and obtains morpheme information included in the first reputation information (S103). Examples of morphological analysis include non-patent literature (Yuji Matsumoto, Kei Kitauchi, Tatsuo Yamashita, Yoshitaka Hirano, Hiroshi Matsuda, Kazuma Takaoka, Masayuki Asahara, “Morphological Analysis System“ Chaya ”version 2.2.1 ”, 2000), for example. The usefulness tag information, the result of the morphological analysis of the first reputation information, and the reputation information are sent to the classification model generation unit 30.

  FIG. 5 is a diagram illustrating an example of morpheme information, which is a result of morpheme analysis. The morpheme information shown in FIG. 5 includes a morpheme surface character string, a morpheme reading, a morpheme basic form, and a morpheme part-of-speech name.

  Again, referring back to FIG. When the classification model generation unit 30 acquires morpheme information, which is the result of morpheme analysis of the usefulness tag information and the first reputation information to which the usefulness tag information is added, the classification of the reputation information is based on these. A classification model used for the above is generated (S104).

  FIG. 6 is a flowchart showing the operation of the classification model generation unit 30. Note that the learning model used in the classification model generation unit 30 is SVM.

  For the morphemes obtained by morpheme analysis, the classification model generation unit 30 has a morpheme occurrence frequency in the first reputation information whose usefulness represented by the usefulness tag information is equal to or greater than a predetermined value, and the usefulness is less than the predetermined value. The appearance frequency in the first reputation information is acquired (S201).

  Next, the classification model generation unit 30 is included in the upper predetermined number of appearance frequencies in the first reputation information in which the usefulness represented by the usefulness tag information is a predetermined value or more among the morphemes obtained by the morpheme analysis, And the morpheme contained in the lower predetermined number of the appearance frequency in the 1st reputation information whose usefulness is less than predetermined value is extracted as useful morpheme (S202).

  Furthermore, the classification model generation unit 30 is included in the upper predetermined number of appearance frequencies in the first reputation information whose usefulness represented by the usefulness tag information is less than a predetermined value among the morphemes obtained by morphological analysis, and The morpheme included in the lower predetermined number of the appearance frequencies in the first reputation information whose usefulness is greater than or equal to a predetermined value is extracted as a non-useful morpheme (S203).

  The classification model generation unit 30 generates a classification model including these useful morphemes and non-useful morphemes as features. The classification model generation method described above is referred to as a screening method. The generated classification model is sent to the classification unit 40 together with reputation information. The classification model generation unit 30 sets all morphemes as targets that can be included in the classification model regardless of the type of part of speech. However, noun morphemes may be excluded. Further, the classification model generation unit 30 may include a syntax pattern such as a morpheme dependency as a feature in the classification model. When using syntactic patterns as features, non-patent literature (Masuichi, Okuma, “Building a Practical Japanese Parsing System Based on Lexical Functional Grammar”, Natural Language Processing Vol.10 No.2, p. 79-109, 2003), LFG or the like described in detail may be used.

  Again, referring back to FIG. Based on the classification model, the classification unit 40 classifies reputation information other than the first reputation information (second reputation information) into useful and unusable information among the reputation information given to the PC 100 (S105). . Specifically, the classification unit 40 considers reputation information that includes useful morphemes that are features in the classification model to be useful, and those that include non-useful morphemes to be not useful. The reputation information to be classified may be all the reputation information given to the PC 100.

  In addition, the classification unit 40 may determine a predetermined classification method, for example, when the number of useful morphemes is larger when one second reputation information includes both useful morphemes and non-useful morphemes. May be useful and may not be useful when there are more non-useful morphemes. Further, the classification unit 40 may be useful or not useful according to a predetermined classification method when neither the useful morpheme nor the non-useful morpheme is included in the second reputation information. Moreover, it may be neither useful nor not useful.

  Hereinafter, experimental results by the inventors of the present invention will be described. In this experiment, a classification model (sPN) generated by applying the screening method described above and a classification model (uPN) generated without applying the screening method were prepared. Furthermore, classification models generated by applying screening methods include those that include parts of noun morphemes (sPN_all) and those that do not (sPN_n), which are considered to be part-of-speech that are easily affected by individual topics. The classification models that were prepared and generated without applying the screening method were prepared to include noun morphemes (uPN_all) and not included (uPN_n), and each was used to classify reputation information.

  In this experiment, 3852 review sentences describing a certain product were used as reputation information for classification. Of these, 1034 review sentences have been determined in advance by a specific user as review sentences that contribute to decision making when considering the purchase of the product, that is, useful reputation information (positive examples). The other 2818 review sentences have been determined by the specific user to be reputation information (negative examples) that are not useful in advance.

  Accuracy, Precision, and Baseline were used as indices representing the accuracy of reputation information classification using a classification model. As shown in FIG. 7 (a), Accuracy is a value obtained by dividing the number of reputation information in which the classification by the system using the classification model and the classification by hand are divided by all the reputation information. Indicates that the reputation information classification using is appropriate. As shown in FIG. 7 (b), Precision divides the number of reputation information classified as useful by humans by the number of reputation information classified as useful by a system using a classification model. The closer to 1, the more appropriate the reputation information classification using the classification model. Baseline is the value obtained by dividing the number of reputation information classified as useful by human resources by all the reputation information, and the reputation information randomly extracted from all reputation information is more useful by human resources. Indicates the probability of being classified.

  FIG. 8 is a diagram showing experimental results. FIG. 8A shows the number of useful morphemes and non-useful morphemes included in the classification model, and when classification is performed by the system using the classification model. FIG. 8B shows the number of useful morphemes and non-useful morphemes included in the classification model, and Precision when classification is performed by the system using the classification model.

  In FIG. 8A, the Accuracy 201 corresponding to the classification model (sPN_all) generated by applying the screening technique and the Accuracy 202 corresponding to the classification model (sPN_n) are the classification models generated without applying the screening technique. The value is larger than Accuracy 203 corresponding to (uPN_all) and Accuracy 204 corresponding to the classification model (uPN_n), indicating that the classification model generated by applying the screening method has higher classification accuracy.

  Further, in FIG. 8B, the Precision 211 corresponding to the classification model (sPN_all) generated by applying the screening technique and the Precision 212 corresponding to the classification model (sPN_n), and generated without applying the screening technique. The Precision 213 corresponding to the classification model (uPN_all) and the Precision 214 corresponding to the classification model (uPN_n) are both larger than Baseline, and classification using these classification models is more effective than random sampling. Is shown. Also, Precision 211 corresponding to the classification model (sPN_all) generated by applying the screening technique and Precision 212 corresponding to the classification model (sPN_n) correspond to the classification model (uPN_all) generated without applying the screening technique. Therefore, the classification model generated by applying the screening technique has higher classification accuracy than the Precision 214 corresponding to the Precision 213 and the classification model (uPN_n).

  As described above, the PC 100 as the reputation information processing system of the present embodiment acquires the usefulness for the specific user for the first reputation information that is a part of the given reputation information, and Perform morpheme analysis, and based on the usefulness for these specific users and the morpheme obtained by morpheme analysis, the morpheme to be included in reputation information with high usefulness and the morpheme to be included in reputation information with low usefulness Is generated, and the second reputation information is classified into those useful for a specific user and those not useful based on the classification model.

  Thus, since the classification model includes morphemes that should be included in reputation information with high usefulness and morphemes that should be included in reputation information with low usefulness, a large amount of reputation information is identified based on the classification model. It can be classified into those that are useful to the user and those that are not useful.

  In the above-described embodiment, when the usefulness tag information according to the setting of the specific user is sequentially added to the reputation information, and the number of reputation information that the usefulness tag information is blown becomes a predetermined value, A classification model may be generated.

  FIG. 9 is a flowchart showing the operation of the PC 100 as a reputation information processing system when usefulness is sequentially added to reputation information.

  The classification model generation unit 30 determines whether or not the number of first reputation information to which usefulness tag information according to the setting of a specific user is added is greater than or equal to the number that enables SVM machine learning. Determine (S301).

  When the number of the first reputation information to which the usefulness tag information according to the setting of the specific user is added is a number capable of machine learning of SVM, the classification model generation unit 30 is added to the first reputation information. A classification model is generated based on the usefulness tag information and morpheme information that is the result of the morpheme analysis of the first reputation information (S302). The specific operation is the same as the operation shown in FIG.

  Next, the classification model generation unit 30 determines whether the ratio of the first reputation information to which the usefulness tag information is manually added to the overall reputation information is n or less which is a predetermined value. (S303).

  When the ratio of reputation information to which usefulness tag information is manually added is greater than n, it is considered that a classification model has been generated according to a sufficient number of first reputation information, and the series of operations ends. . On the other hand, when the ratio of reputation information to which usefulness tag information is manually added is n or less, the classification unit 40 extracts useful information from reputation information other than the first reputation information based on the classification model ( S304). The specific operation is the same as the operation in S105 of FIG.

  Next, the classification unit 30 displays reputation information extracted as useful as a target of addition of usefulness tag information on the monitor 106, and prompts a specific user to set usefulness (S305).

  The specific user operates the operation unit 105 to set the usefulness for the displayed reputation information. The usefulness acquisition unit 10 adds usefulness tag information representing the usefulness set by the specific user to the corresponding reputation information (S306). Furthermore, the usefulness acquisition unit 10 determines whether or not the number of pieces of reputation information to which usefulness tag information is newly added is equal to or more than a predetermined value m (S307).

  When the number of the second reputation information to which the usefulness tag information is newly added is m or more, the morpheme analysis is performed by the morpheme analysis unit 20 as the new first reputation information, and the classification model The operations after the generation of the classification model (S302) by the generation unit 30 are repeated. When the number of reputation information with newly added usefulness tag information is less than m, the operations after the extraction of useful information from the second reputation information by the classification unit 40 (S304) are repeated.

  In S301, when the number of first reputation information to which the usefulness tag information is manually added is less than the number capable of SVM machine learning, the usefulness obtaining unit 10, the classification model generating unit 30, Or the classification | category part 40 extracts arbitrary things from the reputation information to which the usefulness tag information is not added manually (S308). Thereafter, the operations after the display of the reputation information extracted as the addition target of the usefulness tag information by the classification unit 30 (S305) are repeated.

  Thus, when the number of reputation information to which usefulness tag information is sequentially added according to the setting of a specific user reaches a predetermined value, by generating a classification model while using the usefulness tag information, It is possible to minimize the acquisition of usefulness for the first reputation information to the minimum necessary for the classification model used for classification of reputation information to be appropriate, and to quickly generate a classification model. In addition, it is possible to reduce the processing burden and the load of the usefulness setting operation by a specific user.

  Further, the specific user may be plural instead of single. FIG. 11 is a diagram illustrating an example of setting usefulness when there are a plurality of specific users. As described above, when the usefulness for a plurality of specific users is set, although there are variations in the setting of the usefulness of each specific user, a large amount based on the classification model generated according to the usefulness Can be classified into those useful for virtual specific users representing a plurality of specific users and those not useful. According to the inventors' experiment, it was confirmed that classification using a classification model corresponding to a plurality of specific users is more effective than random sampling.

  In the above-described embodiment, the reputation information processing system is configured by a single PC 100, but the usefulness acquisition unit 10, the morpheme analysis unit 20, and the classification model generation unit 30 are configured in one PC, and the classification model After the generation, the classification unit 40 is configured in the other PC connected to the one PC through the communication network so that the reputation information is classified using the classification model generated by the one PC. Good.

  As described above, the reputation information processing system, the reputation information processing method, and the reputation information processing program according to the present invention can obtain reputation information useful for a specific user, and are useful as a reputation information processing system and the like. is there.

It is a figure which shows the hardware constitutions of a reputation information processing system. It is a figure which shows the functional block diagram of a reputation information processing system. It is a flowchart which shows operation | movement of a reputation information processing system. It is a figure which shows an example of the usefulness setting for a specific user. It is a figure which shows an example of the information of a morpheme which is a result of a morpheme analysis. It is a flowchart which shows the operation | movement of classification model generation. It is a figure which shows the parameter | index showing the classification | category precision of reputation information using a classification model. It is a figure which shows the result of the classification experiment of reputation information using a classification model. It is a flowchart which shows operation | movement of the reputation information processing system when usefulness is added to reputation information sequentially. It is a figure which shows an example of the setting of usefulness in case there exist multiple specific users.

Explanation of symbols

10 Usefulness Acquisition Unit 20 Morphological Analysis Unit 30 Classification Model Generation Unit 40 Classification Unit 100 PC
102 Memory 103 HDD
105 Operation unit 106 Monitor 107 Internal bus

Claims (13)

Reputation information representing the evaluation of the object, and acquisition means for acquiring the usefulness of the reputation information for a specific user;
Analysis means for performing at least morpheme analysis among the syntactic and semantic analysis on the acquired reputation information;
A reputation information processing system comprising: generating means for generating a classification model composed of morphemes obtained by the morpheme analysis of reputation information corresponding to the acquired usefulness for a specific user.   The reputation information processing system according to claim 1, wherein the reputation information is information in a sentence unit, and the usefulness is given to each sentence.   The generation means includes a set of morphemes that are included in the upper predetermined number of appearance frequencies in reputation information whose usefulness is greater than or equal to a predetermined value among morphemes obtained by the morpheme analysis, and the usefulness is less than a predetermined value. The reputation information processing system according to claim 1 or 2, wherein a classification model having a set of morphemes included in a top predetermined number of appearance frequencies in reputation information is generated.   The generating means includes, among the morphemes obtained by the morpheme analysis, reputation information that is included in an upper predetermined number of appearance frequencies in reputation information whose usefulness is greater than or equal to a predetermined value and whose usefulness is less than a predetermined value A set of morphemes included in the lower predetermined number of appearance frequencies in the above and reputation information included in the upper predetermined number of appearance frequencies in the reputation information whose usefulness is less than a predetermined value and whose usefulness is greater than or equal to the predetermined value The reputation information processing system according to claim 3, wherein a classification model having a set of morphemes included in a lower predetermined number of appearance frequencies in is generated.   The said acquisition means acquires the usefulness for the specific user of this reputation information according to the setting of the usefulness with respect to the said reputation information by the said specific user, The Claim 1 thru | or 4 characterized by the above-mentioned. Reputation information processing system.   The reputation information processing system according to claim 1, wherein the obtaining unit obtains a usefulness level for the specific user for a part of the reputation information.   The reputation information processing according to any one of claims 1 to 6, further comprising classification means for classifying the given reputation information according to the usefulness for the specific user based on the classification model. system.   The acquisition unit continues to acquire the usefulness of the reputation information for the specific user until the number of the reputation information from which the usefulness for the specific user is acquired reaches a first predetermined value. The reputation information processing system according to any one of claims 1 to 7. The acquisition unit is configured to update the reputation information according to the setting by the specific user until the number of reputation information whose usefulness for the specific user is acquired according to the setting by the specific user reaches a second predetermined value. Continue to acquire usefulness for specific users,
The generation means sets the usefulness for the acquired specific user when the number of reputation information acquired for the specific user according to the setting by the specific user reaches a second predetermined value. The reputation information processing system according to claim 8, wherein a classification model composed of morphemes obtained by the morphological analysis of corresponding reputation information is generated. Reputation information representing the evaluation of the object, and an acquisition step of obtaining the usefulness of the reputation information for a specific user;
An analysis step of performing at least a morphological analysis of the syntactic and semantic analysis on the acquired reputation information;
And a generation step of generating a classification model composed of morphemes obtained by the morpheme analysis of the reputation information corresponding to the acquired usefulness for a specific user.   The generation step includes a set of morphemes included in a higher predetermined number of appearance frequencies in reputation information whose usefulness is equal to or greater than a predetermined value among morphemes obtained by the morpheme analysis, and the usefulness is less than a predetermined value. The processing method according to claim 10, further comprising: generating a classification model having a set of morphemes included in an upper predetermined number of appearance frequencies in reputation information. Reputation information representing the evaluation of the object, and an acquisition step of obtaining the usefulness of the reputation information for a specific user;
An analysis step of performing at least a morphological analysis of the syntactic and semantic analysis on the acquired reputation information;
And a generation step of generating a classification model composed of morphemes obtained by the morpheme analysis of reputation information corresponding to the obtained usefulness for a specific user.   The generation step includes a set of morphemes included in a higher predetermined number of appearance frequencies in reputation information whose usefulness is equal to or greater than a predetermined value among morphemes obtained by the morpheme analysis, and the usefulness is less than a predetermined value. 13. The program according to claim 12, wherein a classification model having a set of morphemes included in a high-order predetermined number of appearance frequencies in reputation information is generated.

Images